Robustness of Causal Links for Precipitation Anomalies and related Variables in Southeastern Africa

Abstract

Southern Africa's society and economy are relying on rain-fed agriculture and energy production, such as hydropower. Therefore, understanding the complex interplay of the climate leading to the variability of precipitation is crucial. In this thesis, causal discovery, which aims to infer the causal structure from a given set of variables, is used to identify the connections between precipitation and its drivers in southern Africa and Madagascar using reanalysis data. The analysis focuses on the rainy season during the austral summer (November-April) for the time period 1979-2014. Despite the highly variable climate system in this region, robust links could be identified with causal discovery for the rainy season using the bootstrapping and the False Discovery Rate control approach, which reduce the number of conflicts and false positives in the causal networks. Precipitation in southeastern Africa is positively influenced by sea surface temperatures in the western Indian Ocean and negatively influenced by a high pressure system located over Botswana. The El Nino-Southern Oscillation has a possible negative influence on the precipitation, in some cases indirectly through the Botswana High. On Madagascar, the precipitation is mainly influenced by the moisture flux of the northeastern monsoon and later in the rainy season by the Mozambique Channel Trough, which deflects the winds towards the island. Overall, the regional variables, such as wind, pressure and sea surface temperatures, drive the precipitation pattern in southern Africa and Madagascar, while little influence is found for the teleconnections. To compare the resulting causal networks between different data sets the F1-score is calculated and a second reanalysis data set is used to test the robustness of the results. In addition, causal networks calculated for three climate models from the High-Resolution Model Intercomparison Project are compared for two resolutions each to analyze the ability of these models to reproduce the links found for the reanalysis data. The highest similarity is found between the causal networks calculated on the basis of all six summer months, between the reanalysis data sets as well as for the model results. Different three-month periods during the austral summer show a higher variety. However, for this analysis no greater similarity was found for higher spatial resolution model runs with reanalysis. Nevertheless, the analysis confirmed the main drivers for precipitation in southern Africa, despite the complex behavior of the climate system. This work provides a basis for further investigation of climate projections and their implications for precipitation in southern Africa in the coming decades.